.claude/skills/video-script-production/references/voice.md

# Spoken voice: taking the machine out of the sentences

Added after the Moser Baer v4 review ("at times it feels full of AI speak: no x, no y, just a simple z"). The fix is a pass at the end of writing, plus budgets in the linter (F rules), not a new way of drafting.

## What this rests on

| Source | What it gives us | Caveat |
|---|---|---|
| Wikipedia, *Signs of AI writing* (WikiProject AI Cleanup) | The catalogue: negative parallelism ("not just X, it's Y"), rule of three, superficial "-ing" riders, vague attributions, promotional tone, "serves as" for "is", vague connection words, section-ending summaries, stock vocabulary, heavy dashes | The page itself says these appear in human editorials, blogs and fiction, that they are potential signs only, and that humans are poor at telling the difference. Use them as a **checklist for flat prose, not a detector.** |
| `blader/humanizer` skill (26 patterns, mark, draft, audit, rewrite) | The useful distinction: patterns that justify an edit on **one** sighting (not-X-but-Y, one-line closers, deep-sounding sayings, staged run-ups like "Here's the thing", arguing with nobody) versus ones that count only when several share a passage (dashes, stacked hedges, passive). Also: changes how text sounds, never what it says; never invents names, numbers or quotes | It targets written prose and wants a voice sample; a script has no sample until we have a recording. |
| `hardikpandya/stop-slop` skill | Structure families worth naming: binary contrasts, negative listing ("It wasn't X. It wasn't Y. It was Z."), dramatic fragmentation, rhetorical setups, false agency, a narrator floating above the scene. A five-axis check: directness, rhythm, trust, authenticity, density | Its blanket bans (no questions starting with wh-, no adverbs, no three-item lists, no dashes) are written for essays. Voice-over uses questions and the occasional list on purpose, so the linter uses budgets. |

Neither skill is evidence that a fix improves retention. The case for it is the plain one: a viewer who hears the same five sentence shapes repeated stops hearing the content. Treat the F rules as a budget on repetition.

## The tells, as they show up in a spoken script (Moser Baer v4 examples)

| Tell | Why it reads as machine | Fix | Before, then after |
|---|---|---|---|
| **Negation pivot** (F1) | It denies something nobody said, then says the thing. Three in a row is the "no X, no Y, just Z" shape. | Say the thing. | "no crash, no scandal, just a margin so thin that a small slip wipes out everything" becomes "From outside it looked fine. The margin was so thin that a 5 percent dip in sales erased a year's profit." |
| **Staged run-up** (F2) | Announces that a point is coming instead of making it. | Cut the announcement; the next sentence is the point. | "Think about what that does. A shrinking business has no date." becomes "A shrinking business has no date." |
| **Script-talk** (F2) | The narrator talks about the script: promises, reminders, naming the idea aloud. | Use the name; skip the ceremony. | "I call that the treadmill, and I'll keep using the word." becomes "Call it the treadmill." |
| **Stock depth** (F3) | Sounds profound, says nothing. | Replace with the fact it gestures at. | "Brilliant is the right word." becomes "By 2007 it was the second-biggest maker of recordable discs in the world." |
| **Triad** (F4) | Three items because three sounds finished. | Use the number of items there are. | "It sells, it shrinks, it protects what is left." becomes "It sells what it can and shrinks to what it can afford." |
| **Punch-line closer on every beat** (F4) | A six-word zinger ending each paragraph makes every paragraph the same shape. | Let some beats end on the hinge question, some on a plain fact. | |
| **Fragment drumroll** (F4) | Short, short, short. Manufactured emphasis. | One complete sentence. | |
| **Written English** (F5) | No contractions, because it was typed, not said. | Read it aloud; contract where you would. The twelve English winner transcripts we have run about 1.5 contractions per 100 words (range 0 to 3.7). | |

## The pass (Step 4b of SKILL.md)

Run it **after** the texture pass (Step 4a), which adds a persona, figures, people and landings. This pass removes the tells those additions may bring ("It was like a", "Think about what that does", a staged aside). Add, then subtract.

1. **Mark.** Run `lint.js`; read the F detail lines. Then read the script aloud once at speaking pace and mark every place you stumble, run out of breath, or would not say it to a friend.
2. **Rewrite by pattern, not by sentence.** For each tell, ask what the sentence is *for*. If it carries a fact, keep the fact and drop the frame. If it only announces, cut it. Where you cut a sentence that also carried a hinge or a care line, move the hinge into the neighbouring sentence and re-check C and D.
3. **Specificity test.** Replace each abstract line (the situation, the pattern, the stakes) with a name, a number, a date, an object or a quote that is already in the fact ledger. Never add a fact that is not in the ledger; the humanizer rule applies: change how it sounds, never what it says.
4. **Keep the honest hedges.** "I couldn't find where that money went" is both the most human line in the script and the one that protects the claim.
5. **Do not over-correct.** A contrast that carries information stays ("In December the company said it was well-funded. The quarter before, it had made a loss."). A question the viewer is actually asking stays. The linter's budgets leave room for these; if a rule fires on a sentence that carries a fact, keep the sentence and note why in the lint log.
6. **Re-run lint**, then the fresh-reader pass with one added question: *which sentences sound written rather than spoken?*

## What the F rules measure

| Rule | Measures | Default budget (`rules.json` > `voice`) |
|---|---|---|
| F1 | negation pivots | warn above 1, fail above 4 |
| F2 | staged run-ups and script-talk | warn above 3, fail above 8 |
| F3 | stock phrases, machine vocabulary, empty intensifiers, lazy extremes | warn above 2, fail above 8; extremes under 6 per 1000 words |
| F4 | fragment runs, triads, short closers, sentence-length variation, repeated openers | fail above 3 fragment runs; the rest warn |
| F5 | contractions, dashes, average sentence length | warn |

Patterns live in `scripts/voice-patterns.json`. Add a pattern when a reader or a new skill names a tell the list misses, and add the sentence that taught it to the table above.

## What it cannot do

It finds the tells that a regex can find. A script can pass every F rule and still read like a template, and a script with a few tells can still sound like a person. The final test is a person (or a fresh reader) hearing it. Detectors are not the goal; the humanizer skill's own page notes that detectors still flag most of its output, and flag plenty of human text too.